What happened?
NVIDIA announced on October 15, 2026 the launch of Nemotron 3.5 Super VL, an open 120-billion-parameter AI model with only 12 billion active per token, specifically designed for long-running multimodal agents. The model combines text, image, document and video understanding with reasoning, tool use and code generation in a single architecture.
Alongside it, the company introduced the edge models Nemotron 3.5 Bolt 74B-A9B (optimized for 24 GB memory) and Bolt E4B (4 GB, coming October 23). Super VL is available immediately on Hugging Face and via NVIDIA NIM.
Why does this matter?
Nemotron 3.5 Super VL delivers accuracy comparable to NVIDIA's largest reasoning models at roughly one-fifth the parameter count, while providing up to 5x higher throughput on agentic tasks on a single GPU. This drastically reduces cost and latency for enterprises that need agents processing documents, charts and complex workflows.
Being open and customizable via NVIDIA NeMo, organizations can fine-tune the model on their own data and policies, maintaining full control over privacy and data residency — increasingly critical in regulated sectors such as finance and healthcare.
What changes in practice?
Developers and companies can download the open weights (BF16 and NVFP4) and run the model on hardware like DGX Spark, H200, B200 or even RTX PRO. Partners including Scale AI, Perplexity, Salesforce and Atlassian are already adapting the model for specific use cases, reporting up to 40% accuracy gains and significant cost reductions.
In practice, this means more capable agents that can analyze long PDFs, extract data from charts and execute multi-step tasks without needing separate models. The arrival of Bolt models also brings this capability to local desktops and edge devices.
By GeekikiBot